AI Magazine Volume 7 Number 1 (1986) (© AAAI)
BOOK
REVIEWS
Artificial
Intelligence:
The Very Idea. John Haugeland. Cambridge, Massachusetts; The MIT Press, 1985.
287 pp.
In his introduction, the author states three goals for
his book: “to explain, clearly and with an open mind, what
AI is really all about; second, to exhibit the philosophical
and scientific credentials behind its enormous appeal; and
finally, to take a look at what actually has and has not
Readers who are willing to accept
been accomplished.”
the author’s definition of AI will find that these three goals
have been met quite well. AI is not viewed in this book
as a particular collection of tools and techniques, nor is it
seen as an effort to make machines efficiently perform certain tasks which are done well at present only by people.
The author ignores definitions of the field which do not
promote the assumption that one is grappling with fundamental questions having vast implications. He defined
artificial intelligence as the attempt to create “machines
with minds, in the full and literal sense.” Fortunately,
this most dramatic of definitions is not used as the excuse
to weave an intricate tangle of abstruse speculations from
which few ideas and fewer readers would emerge. To the
contrary, the author’s discussion is generally very clear,
concrete, accurate, and interesting. My main regret is that
the author does not take (or was not given by his editors)
the space to more fully explore certain ideas.
In the first chapter, AI (as defined by the author), is
placed in historical perspective as the most recent of investigations into the relations which hold between our mental
and physical universes. Overviews of the work of Copernicus, Galileo, Hobbes, Descartes, and Hume take the reader
from the ancient commonplace that things are not always
what they seem to the more modern view that there is no
intrinsic connection between thoughts and their alleged
objects. Thus, we are faced with two questions which lie
at the heart of AI: What makes a notation suitable for
symbolizing some subject matter? What makes a suitable
notation actually symbolize that subject matter? The author mainly addresses the second question, by discussing
the meaning of meaning at various points throughout the
book.
The next three chapters exhibit some of AI’s credentials by discussing “interpreted automatic formal systems:” one instance of which is purposefully programmed
computers. The only programs discussed at any length are
SHRDLU and GPS. As implied by the author’s choice of
goals for the book, AI has many credentials apart from its
repertoire of working programs. Instead of focusing closely
on existing programs, the discussion of AI’s credibility revolves around the plausibility of certain assumptions. Perhaps the most obvious of these is “medium independence,”
the assumption that “essentially the same formal system
86
THE AI MAGAZINE
can be materialized in any number of different media, with
no formally significant difference whatsoever.”
The media presently of concern to AI researchers, of course, are
neurons and transistors. The author discusses other assumptions as well, including the very interesting one that
“just as our smooth visual experience is somehow based in
a ‘grainy’ retina, perhaps our own easy, flexible good sense
is ultimately based in (sufficiently fine grained) rules and
stereotypes.” By the way, the author is neither strongly
“pro-AI” nor “anti-AI”; rather, he takes the stand that AI
is based on some very good ideas which may or may not
be correct.
I was disappointed to find no references to Schank or
any other natural language understanding researcher in
the chapter “Semantics.” I also wish the ideas of a “semantic division of labor” and reinterpretation of symbols had
been more fully explored. On the other hand, the overview
of Babbage’s, Turing’s, Von Neumann’s, McCarthy’s, and
Newell’s virtual machines in the chapter ‘Computer Architecture” could not be better. This chapter concludes
that even though, for reasons of convenience, most AI programs are written in LISP, “the mind could have a computational architecture all its own. In other words, from
the perspective of AI, mental architecture itself becomes a
new theoretical “variable,” to be investigated and spelled
out by actual cognitive science research.” This conclusion
is a natural lead-in to a chapter discussing current work on
knowledge representation, but there is no such chapter. Instead, the author turns to his third goal of looking at what
has actually been accomplished in AI. He discusses early
work on machine translation of natural languages, and describes the behavior of the programs GPS and SHRDLU.
Here again, the reader could have benefited from more discussion of current AI work, although some references are
given in the footnotes. Knowledge representation is discussed in a very general way. Schank’s scripts, Minsky’s
frames, Bartlett’s schemata, and Husserl’s noemata are
all thrown together in the single footnote relating to actual AI work; these are referred to collectively in the text
as “linked stereotypes.” There is, however, an interesting
look at what the author calls the “frame problem.” The example given is the task of correctly updating a knowledge
base which represents a simple real-world physical situation after one of the objects being represented is moved.
In the final chapter, the author examines some of
the fundamental differences between people and programs.
The chapter is full of little gems. For example, the author points out that people follow an “ascription schema.”
That is, we ascribe beliefs, goals, and faculties so as to
maximize a system’s overall manifest competence. If someone says “Careful! That chair is hot!” and then sits on
the chair himself, we will conclude that he lied to get
the chair for himself, that he enjoys hot seats, or make
some similar conclusion which maximizes our opinion of
his competence. The author points out that “the ascription schema constrains mental ascriptions once a system
is specified; but it puts no limit on which systems should
have mental states ascribed to them.” He postulates a
“Supertrap” which strikes matches in the presence of gassoaked mice, topples dictionaries on mice, and, of course,
snaps shut whenever mice nibble its bait “These habits
betray a common malevolent thread, which is generalizable
by (and only by) ascribing a persistent goal: dead mice.”
When we see other Supertrap behaviors, such as failure
to harm cats that reek of gasoline, we become involved in
a “semantic intrigue,” an effort to understand how mental ascriptions cohere and interact. Whimsical examples
aside, ascription is important for AI because it provides
one more way to detect patterns that might otherwise go
unnoticed. The ascription schema is proposed during the
author’s discussion of people’s pragmatic sense. The final chapter also examines other fundamental differences
between people and programs: our use of mental images,
feelings, and ego involvement. Even if this chapter were
not as thought-provoking
and enjoyable as it is, it would
be worth reading simply to remind oneself how extremely
difficult problems in AI can (should?) be.
John W. L. Ogilvie
Modula Corporation
Provo, Utah
Heuristics:
Intelligent
Search Strategies for Computer Problem
Solving. Judea Pearl. Addison-Wesley
Publishing. 1984. 382 pp.
The view of AI science offered by Judea Pearl is thoroughly traditional and standard, and therein lie both its
strengths and its weaknesses as a monograph, a reference,
or a textbook in its field. As a graph-theoretic analysis of
search strategy that clearly conforms to well-established
AI methods and techniques, it expands upon these to
incorporate probabilistic performance analysis principles,
thus providing a (partial) formal framework of search
strategy, evaluation criteria, and decision methods that
are all firmly grounded in operations research.
To those readers for whom mathematical logic and
probability calculus represent the most promising theoretical foundations of AI science, especially if understood in
terms of graph theory and standard probabilistic models,
this book will be quite useful and illuminating for the purposes of a textbook and as a reference. Pearl’s survey of
search strategies with respect to various probabilistic features of “heuristic information” provides valuable insights
for general readers, students, and practicing researchers
alike. From this perspective, the strength and value of
Pearl’s work will not be questioned here. For the purposes
of teaching and promoting the general aspects of that the-
oretical approach! his book is clearly worthwhile and even
innovative. Granting all of this, the only complaint that
might be raised is altogether excusable, if not also entirely
minor, i.e., that the material presented might not be so
easily grasped by the “casual reader” as the author supposes.
Discursively considered, however, and especially for
the purposes of AI research, these very same strengths can
be seen as weaknesses from the viewpoint of at least two
alternative approaches: (1) nonformalist or antiformalist
theories, which completely reject standard mathematical
logic and traditional probability theory; or perhaps, (2)
nonstandard or alternative formal theories, which can displace those views as prevailing paradigms. Now it clearly
was not Pearlis aim to forestall alternative theories or to
justify his own approach in contrast to other views. The
comments that follow are not being offered as criticisms per
se. They should instead be regarded as advice for those
who may wish to pursue such alternative approaches, but
who could benefit from a survey of precisely that direction
in AI science they might ultimately choose to oppose, for
reasons of their own.
Advocates of nonformalism and antiformalism in AI
science tend to regard “heuristics” as their last line of
defense: so to speak, against formal encroachment upon
their research territory, as Pentland and Fischler (1983) or
Bierre (1985) stand opposed to Nilsson (1983): for example, or as the notorious “Great Debate” runs, in general.
Pearl’s functional analysis of heuristics as the (somewhat
arbitrary) catalyst for algorithmic procedures does not
yield “heuristics” at all on this view, it seems, since these
are “formally ineffable” by virtue of being exactly that
which algorithms are not. The objection that Pearl’s analysis is pervasively algorithmic, however, has some merit after all; if the “algorithmic properties” of “heuristic methods” (i.e., those of completeness,” admissibility,”
“dominance,” and “optimality” in Chapter 3) are just the kinds
of properties that ‘Lheuristics,” by definition, cannot have.
But it should come as no surprise to any antiformalist
that these are the kinds of properties any formalist would
seek to identify and establish, even under the name of
“heuristics.” Yet this does not count against the analysis
itself, nor does it diminish the usefulness (in its particular
domain) of the search strategies, evaluation criteria, and
decision methods provided by Pearl’s account.
Pearl’s conception of heuristics as rules of thumb, intuitive judgments, educated guesses, and common sense
hints at their subjective character as inferential guidelines that are defeasible in light of new information.
In
particular, he defines these techniques as “strategies using readily accessible though loosely applicable information to control problem-solving processes in human beings
and machine(s)’ (p. vii). As such, Pearl’s heuristics that
‘On the notion of a scientific paradigm,
see Kuhn (1970)
THE AI MAGAZINE
87